llama.cpp 1.0 Locks In a Stable C API

llama.cpp, the open-source engine behind much of today's local AI software, reached version 1.0.0 on Nov. 13, together with ggml, the tensor library it runs on. The ggml-org team at Hugging Face said the milestone, about 3.5 years after the project began, comes with a formal promise: from now on, changes that break the libraries' public interfaces will show up as major-version bumps.

Key takeaways

• Both libraries, libllama and libggml, moved to semantic versioning with v1.0.0, so developers can tell from the version number whether an update may break their code.
• The release removes 35 deprecated functions and adds FP8 KV caches, a host-memory cache for mixture-of-experts models and new model support.
• llama.cpp now supports more than 150 architecture variants and ships through eight distribution channels, from Docker to Winget.

What llama.cpp 1.0 promises developers

Until now, llama.cpp was versioned by build numbers, and the last numbered releases were llama.cpp v0.9.0 and ggml v0.29.0. In its announcement, the team wrote that "many downstream applications already depend on our work" and that "having a flexible but stable API is paramount."

The team said both libraries will reflect API and ABI incompatibilities this way and that it will become "stricter about making those boundaries visible" as adoption grows.

What changed in the v1.0.0 release notes

The llama.cpp release notes show a cleanup before the stability pledge. Pull request 30565 removes llama_set_warmup, 35 deprecated functions and the defrag_thold option, and moves the NextN speculative-decoding API into the main llama.h header.

New features include FP8 KV caches and a host-memory expert cache for mixture-of-experts models with a configurable CPU share. Added models include Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0, plus drama-3 text-to-speech and two vision models.

The built-in llama-server gets fixes for CORS, DNS rebinding and server-side request forgery, cancels MCP tool calls when a client disconnects, and adds an OpenAI-compatible /v1/audio/speech endpoint. Its web interface adds named chat groups and a live-preview markdown editor.

The ggml 1.0.0 notes list FP8 support, configurable YaRN truncation and fused mixture-of-experts routing on Apple's Metal for DeepSeek-V4 and V4.1, the model family AI Newsway covered earlier.

Why a stable llama.cpp matters for local AI

The project says it supports more than 150 architecture variants and all mainstream speculative decoding methods. Its CI runs on self-hosted machines with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Builds are distributed as pre-release artifacts for Linux, macOS, Windows, Android and iOS, and through Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework. The team also points users to its Llama desktop apps for macOS and Windows.

What's next

The real test is the next breaking change. If the team holds to its rules, any future removal from the public C API will arrive as v2.0.0 rather than inside a routine update. Developers linking against llama.cpp should first check the functions removed in pull request 30565.

FAQ

Is llama.cpp an AI model?
No. It is an inference engine that runs models, mostly in the GGUF format, on local hardware.

Do I need to change my code to upgrade to 1.0.0?
Only if your software calls one of the removed functions, such as llama_set_warmup or the deprecated APIs listed in pull request 30565.

What does a minor version bump mean now?
A backward-compatible addition, such as a new feature, backend update or model support.

References

• llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem — Hugging Face (Nov. 13, 2026)
• Release v1.0.0 · ggml-org/llama.cpp — GitHub (Nov. 13, 2026)
• Release v1.0.0 · ggml-org/ggml — GitHub (Nov. 13, 2026)